Industries / Finance and business services / Translation services
Legal and medical translation
Translating and interpreting where a mistake can decide a case or harm a patient: in courts and police stations, in the text of laws and patents, and between clinicians and patients.
7
Milestones achieved, from 1958 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1958
Law in every language
The European Economic Community's first regulation, published on 6 October 1958, requires regulations and other documents of general application to be drafted in the official languages and the Official Journal, where EU law is published, to appear in all of them. Amended at each enlargement, it has listed 24 official languages since 2013. The Court of Justice holds all language versions equally authentic and first ruled on a conflict between them in 1967, in the Van der Vecht case, by the account of Stefaan van der Jeught of the Vrije Universiteit Brussel.
LessonWhen every version is authentic, a mistranslation in one becomes a legal question that only the court can settle by reading the others.
Credit
Council of the European Economic Community; Court of Justice of the European Union
1978
Certified interpreters for federal courts
The US Court Interpreters Act of 28 October 1978 told the Administrative Office of the US Courts to set the qualifications of certified court interpreters, keep a master list of them and set their fees, and told judges to use the most available certified interpreter in actions brought by the United States. The federal courts now sort interpreters into certified, professionally qualified and language-skilled, and their fee page in November 2026 listed $617 for a full day of a certified interpreter's work.
Credit
US Congress; Administrative Office of the US Courts
1984
A word worth $71 million
An unconscious 18-year-old was brought to an emergency department, and his mother and girlfriend used the Spanish word "intoxicado", by which they meant he was nauseated. With no one there who spoke Spanish, staff took it to mean intoxicated and treated him for a drug overdose. Nearly three days later a test showed two blood clots in his brain from a break in an artery that had been defective since birth, and he was left quadriplegic. His settlement topped $71 million; the magazine Medical Economics reported the case in 1984 under the headline "A misinterpreted word worth $71 million".
LessonA single word taken at face value, with no trained interpreter present, can decide a diagnosis.
Credit
reported by Medical Economics; retold by the AMA's journal Virtual Mentor in 2007
2003
Counting interpreters' errors
Glenn Flores and six colleagues recorded 13 pediatric visits at a hospital outpatient clinic that used Spanish interpreters and found 396 interpreting errors, an average of 31 per visit; 63% could have had clinical consequences. The most common error was leaving something out (52%). Errors by ad hoc interpreters, such as nurses, social workers and an 11-year-old sibling, were more likely to matter clinically than those by the hospital's own interpreters (77% against 53%). One error told a mother to put amoxicillin, an antibiotic, in both of her child's ears to treat an ear infection.
LessonThe study gave hospitals a measured case for paying trained interpreters instead of using whoever is at hand.
Credit
Glenn Flores and colleagues, published in Pediatrics
2010
A right to interpretation
EU Directive 2010/64/EU of 20 October 2010 gives anyone suspected or accused in criminal proceedings in the EU a right to interpretation and to translation of essential documents, of a quality "sufficient to safeguard the fairness of the proceedings", and a way to complain when it is not. Member states had until 27 October 2013 to make it law and were asked to try to keep registers of qualified independent translators and interpreters.
Credit
European Parliament and Council of the European Union
2024
Machine translation needs review
A rule under Section 1557 of the US Affordable Care Act, published by the Department of Health and Human Services on 6 May 2024 and in effect from 5 July 2024, says that when a health program uses machine translation for text that is critical to a patient's rights or benefits, or where accuracy is essential, a qualified human translator must review it.
Credit
HHS Office for Civil Rights
2025
Federal language order revoked
Executive Order 13166 of August 2000 told every US federal agency to plan how people with limited English could use its services, and to make sure that organizations it funds did the same. On 1 March 2025 Executive Order 14224 named English the official language of the United States, revoked the 2000 order and told the Attorney General to withdraw guidance issued under it, while saying agencies need not stop offering services in other languages.
LessonLanguage access that rests on an executive order can be undone by the next one; the 2024 health rule above sits in regulation instead.
Credit
the White House, in orders of 11 August 2000 and 1 March 2025
Watch for
An in-house AI interpreter
A team that built its own app, LingualAI, tested it against certified medical interpreters on scripted English-Spanish ear, nose and throat visits, with bilingual clinicians rating anonymized audio. On a 5-point scale it came within the preset margin of the humans on correct medical terms and on keeping the meaning, but not on clarity, and none of the voice measures (fluency, rhythm, pace) met it; overall ratings favored the interpreters. The authors suggest using it as an aid when interpreters are not available.
Achieved whena team that did not build the app repeats the test in more languages and on unscripted speech and gets the same result on meaning and terms.
Claimed by
the app's developers, Uday Singh, Martin Citardi and colleagues, in npj Health Systems, May 2026
Open
01
Find divergent EU language versions
A tool that lines up all 24 language versions of an EU regulation or directive sentence by sentence and flags places where one version says something different from the others: a missing negation, a different number or date, a wrong cross-reference, a term that changes the scope. Test it against corrigenda, the corrections the EU publishes after a text has appeared, which give a ready list of known errors.
The briefCredit · unclaimed02
Check numbers across medicine leaflets
A checker that takes the official product information of medicines authorized for the whole EU, which the European Medicines Agency (EMA) publishes in the EU languages and in Icelandic and Norwegian, and compares every number with a unit across language versions: doses, intervals between doses, maximum daily doses, storage temperatures and shelf lives. Each mismatch is then confirmed or rejected by people who read that language.
The briefCredit · unclaimed
Open 01 · the brief
Find divergent EU language versions
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that lines up all 24 language versions of an EU regulation or directive sentence by sentence and flags places where one version says something different from the others: a missing negation, a different number or date, a wrong cross-reference, a term that changes the scope. Test it against corrigenda, the corrections the EU publishes after a text has appeared, which give a ready list of known errors.
Why it is still open
Stefaan van der Jeught, writing in May 2025, notes that "one or more language versions may therefore prove to be inaccurate", that Dutch judges, in an earlier study of his, used no systematic approach to spot divergences, and that AI could detect them both while laws are drafted and afterwards, but he describes no such system in use. The JRC-Acquis, the EU's aligned corpus of its laws, was built for language research, not as a checker. The nearest benchmark, RegTrack from the Technical University of Munich (ACL 2026), labels changes between successive versions of EU regulations, not differences between languages. This research found no tested detector and no labeled set of corrigenda.
languageandlaw.eujoint-research-centre.ec.europa.euaclanthology.org
Built with what exists
EU law and its corrigenda from the Publications Office's open data service, by CELEX number (a corrigendum adds "R(01)", "R(02)" and so on to the act's number); the paragraph ids EUR-Lex gives every language version of an act, which line most of the 24 versions up without a separate alignment step, as the eurlex-corpus project (August 2026) showed for eight financial acts; an AI model given the aligned sentences, such as Claude, Gemini or the open-weight Mistral Large 4 (see the Note "A trillion open parameters"); two bilingual lawyer-linguists to check flags.
First step
Pull the 2024 corrigenda to EU regulations, read 20 of them against the original texts (for example the second corrigendum to the GDPR), and sort the errors into types.
Done looks like
A published, labeled set of at least 300 corrigenda from 2015 to 2025, each with the language and passage it corrects; the tool run on the texts as first published; the share of corrected passages it flags, and the share of its flags that the two checkers confirm as real divergences.
Open 02 · the brief
Check numbers across medicine leaflets
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that takes the official product information of medicines authorized for the whole EU, which the European Medicines Agency (EMA) publishes in the EU languages and in Icelandic and Norwegian, and compares every number with a unit across language versions: doses, intervals between doses, maximum daily doses, storage temperatures and shelf lives. Each mismatch is then confirmed or rejected by people who read that language.
Why it is still open
EMA's procedure (revision 6, February 2025) has national authorities review each product's translations inside a fixed timetable, product by product. Nicola Pelizzari (2024) found no protocol for testing translated patient leaflets with users. This research found no public audit that counts mismatches between language versions across many medicines.
Built with what exists
EMA's product information PDFs, one per language at addresses of the form ema.europa.eu/<language>/documents/product-information/<medicine>-epar-product-information_<language>.pdf (25 language versions for the cancer drug Keytruda, for example); pdftotext or PyMuPDF to extract text; a script that finds numbers with units; an AI model to match sections across languages; native-speaking pharmacists or medical translators to confirm mismatches.
First step
Download one medicine's product information in five languages, extract the numbers from the dosing section, and compare them with the English by hand.
Done looks like
200 medicines checked in all their language versions; every flagged mismatch confirmed or rejected by two reviewers; the rate of confirmed mismatches per 10,000 numbers checked, by language and by section, published with the code; confirmed mismatches sent to EMA.